Seedance 2.5

The most capable clip model in the catalog and the one to reach for when a scene has several characters who all have to stay themselves. Thirty reference images is not a bigger seven. It is a whole cast sheet, and combined with ten reference videos and ten reference audios it is enough to specify who is in a shot, how they move and what it sounds like in one call.

One thing to settle before you build that cast: a photoreal human face is refused unless Seedream made it, and a photograph of a real person or a photoreal character from another image model fails the same check. Drawn characters pass whatever drew them.

It also runs the longest single take here: up to thirty seconds, against the four-to-eight most of the catalog offers. That matters more than it sounds, because the usual way to exceed a model's ceiling is to chain takes through their final frames, and every join is a place continuity can break.

The price needs watching. A five-second 720p take is 96.5 credits from a prompt or image, but 403.5 if you pass a reference *video*, because video-in meters at roughly four times video-out. That multiplier applies at 480p too. Check what you are attaching before you press, because the difference between the two is not a rounding error.